Using Big Data Analytics to “Back Engineer” Protein Conformational Selection Mechanisms

Abstract

In the living cells, proteins bind small molecules (or “ligands”) through a “conformational selection” mechanism, where a subset of protein structures are capable of binding the small molecules well while most other protein structures are not capable of such binding. The present work uses machine learning approaches to identify, in a very large amount of protein:ligand complexes, what protein properties are associated with their capacity to bind small molecules. In order to do so, we calculate 40 physicochemical properties on about 1.5 millions of protein conformations: ligand and protein conformations. This work describes a machine learning approach to identify the unique physico-chemical descriptors of a protein that maximize the prediction rate of potential protein molecular conformations for the test case proteins ADORA2A (Adenosine A2a Receptor), ADRB2 (Adrenoceptor Beta 2) and OPRK1 (Opioid Receptor Kappa 1). We find adequate machine learning techniques can increase by an order of magnitude the identification of “binding protein conformations” in an otherwise very large ensemble of protein conformations, compared to random selection of protein conformations. This opens the door to the systematic identification of such “binding conformations” for proteins and provides a big data approach to the conformational selection mechanism.

Big Data analytics for improved prediction of ligand binding and conformational selection

“…As described in our previous work ( Gupta et al, 2022 ), Molecular Dynamics (MD) simulations of four proteins, namely, ADORA2A (Adenosine A2a Receptor), ADRB2 (Adrenoceptor Beta 2), OPRD1 (Delta Opioid Receptor) and OPRK1 (Opioid Receptor Kappa 1) were used to study the efficacy of our proposed method. The conformations of these four proteins have been well-studied, and the protein conformations that: a) will bind to ligands (binding conformations) and b) will not bind to ligands (non-binding conformations), are known and have been previously documented and published ( Evangelista et al, 2016 ).…”

“…Tables describing the protein attributes/features/descriptors for ADORA2A, ADRB2, OPRD1, and OPRK1 datasets can be found in our previous work ( Gupta et al, 2022 ). ADRB2 and OPRD1 have one additional feature - pro_pl_seq (Sequence based pI) in comparison to ADORA2A and OPRK1.…”

“…ADRB2 and OPRD1 have one additional feature - pro_pl_seq (Sequence based pI) in comparison to ADORA2A and OPRK1. The molecular descriptors were calculated using the protein descriptors from the program MOE ( Akondi et al, 2019 ; Chemical Computing Group, 2019 ; Gupta et al, 2022 ; Sripriya Akondi et al, 2022 ).…”

“…Big Data analytics provides an efficient approach to analyzing such a large amount of data, and also addresses the class imbalance problem ( Abd Elrahman and Abraham, 2013 ), which is a result of imbalanced groups or sub-categories present in the data, where the majority class or larger group of data consists of non-binding protein conformations and it overshadows the minority class or smaller data group, which comprises the data-of-interest i.e., the binding protein conformations. In our prior work ( Akondi et al, 2019 ; Gupta et al, 2022 ; Sripriya Akondi et al, 2022 ), a novel two-stage sampling-based classifier framework was proposed with the primary goal of addressing the class imbalance problem and maximizing the detection of potential binding protein conformations as conventional machine learning (ML) algorithms are ill-equipped to deal with the issue of class imbalance during the data-learning phase. This paper extends on our previous work by presenting additional improvements to our two-stage sampling-based classification approach ( Gupta et al, 2022 ) using deep learning techniques and four different feature selection methods in conjunction with an Enrichment ratio framework.…”

“…In our prior work ( Akondi et al, 2019 ; Gupta et al, 2022 ; Sripriya Akondi et al, 2022 ), a novel two-stage sampling-based classifier framework was proposed with the primary goal of addressing the class imbalance problem and maximizing the detection of potential binding protein conformations as conventional machine learning (ML) algorithms are ill-equipped to deal with the issue of class imbalance during the data-learning phase. This paper extends on our previous work by presenting additional improvements to our two-stage sampling-based classification approach ( Gupta et al, 2022 ) using deep learning techniques and four different feature selection methods in conjunction with an Enrichment ratio framework.…”